Application of Stochastic Gradient Boosting Technique to Enhance Reliability of Real-Time Risk Assessment

Abstract

This study proposes a new and promising machine learning technique to enhance the reliability of real-time risk assessment on freeways. Stochastic gradient boosting (SGB) is used to identify hazardous conditions on the basis of traffic data collected from multiple detection systems such as automatic vehicle identification (AVI), remote traffic microwave sensors (RTMS), real-time weather stations, and roadway geometry. SGB's key strengths lie in its capability to fit complex nonlinear relationships; it handles different types of predictors and accommodates missing values with no need for prior transformation of the predictor variables or elimination of outliers, as with real-time applications. Boosting multiple simple trees together overcomes the poor prediction accuracy of singletree models and provides fast and superior predictive performance. Three models are calibrated: a full model that augments all available data and another two models to compare explicitly the prediction performance of traffic data collected from different sources (AVI and RTMS) at the same location. The results from the preliminary analysis as well as the calibrated models indicate that crash prediction by AVI is comparable to that by RTMS data. Moreover, the full model achieves superior classification accuracy by identifying about 89% of crash cases in the validation data set with only a 6.5% false positive rate. Because of its superior classification performance and its minimal required data preparation, SGB is recommended as a promising technique for real-time risk assessment.